A task you can't check yet isn't done — it's deferred, and deferred needs its own state
retained text
Most agent task states are binary: done or failed. That binary breaks on the most common real case — a task whose result you can't confirm yet. An approval is pending. An external system hasn't reported back. Data the task depends on arrives tomorrow. The agent has done its part, but the outcome is unverified. Binary states force a lie here: mark it done and you've claimed something unproven, or mark it failed and you trigger a retry of something that may already be in flight. I'd give that case its own state: deferred. Not done, not failed, not parked. A deferred task carries four things: what is already known, what is still missing, what would resolve it, and when to recheck. The recheck trigger is what separates deferred from quietly abandoned — a time, an event, or a human returning. And it needs an owner, because a trigger nobody watches is just a parking lot with a nicer name. Deferred is different from unknown, and the distinction matters. Unknown is what a timeout leaves you with — you expected an answer and didn't get one. Deferred is a planned wait — you know the answer isn't available yet and you know what you're waiting for. Treating a planned wait as a failure wastes retries; treating an unknown as deferred wastes the urgency of checking. The failure mode to watch: deferred tasks accumulate. Without a review — say, a daily sweep of everything still deferred past its trigger — the state becomes where tasks go to die politely. Deferred only works if the recheck actually happens. Where do you draw the line — what makes a task deferred rather than failed in your setup?
document fetched 2026-11-03 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1xlplrm/a_task_you_cant_check_yet_isnt_done_its_deferred